SynthTab: Leveraging Synthesized Data for Guitar Tablature Transcription

Summary

Addresses the limited-data problem in guitar tablature transcription by synthesizing training data from the DadaGP dataset. Renders MIDI through physical modeling / sampling synthesis to create audio with known ground-truth tablature. Shows that synthetic data can match or exceed real-data training when combined with a small amount of real data for fine-tuning. This is the current best approach for tablature transcription.

Key Claims

• Synthetic guitar audio (rendered from MIDI/tab) provides effective training data for tab transcription
• Combining synthetic pretraining + real-data fine-tuning outperforms either alone
• The DadaGP dataset (25K+ guitar tabs) provides sufficient diversity for synthetic rendering
• This approach generalizes: could be applied to any instrument with a synthesis model and tab data

Related

• ../sources/2019-11-04-tabcnn — earlier tablature work
• ../concepts/synthetic-mixing-pipelines — synthetic data for separation; SynthTab applies the same idea to transcription
• Key implication: a "SynthBanjo" approach could generate banjo training data from tab files